2 * Copyright © 2010 Intel Corporation
3 * Copyright © 2014-2017 Broadcom
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 * The basic model of the list scheduler is to take a basic block, compute a
29 * DAG of the dependencies, and make a list of the DAG heads. Heuristically
30 * pick a DAG head, then put all the children that are now DAG heads into the
31 * list of things to schedule.
33 * The goal of scheduling here is to pack pairs of operations together in a
34 * single QPU instruction.
37 #include "qpu/qpu_disasm.h"
38 #include "v3d_compiler.h"
39 #include "util/ralloc.h"
43 struct schedule_node_child
;
45 struct schedule_node
{
46 struct list_head link
;
48 struct schedule_node_child
*children
;
50 uint32_t child_array_size
;
51 uint32_t parent_count
;
53 /* Longest cycles + instruction_latency() of any parent of this node. */
54 uint32_t unblocked_time
;
57 * Minimum number of cycles from scheduling this instruction until the
58 * end of the program, based on the slowest dependency chain through
64 * cycles between this instruction being scheduled and when its result
70 struct schedule_node_child
{
71 struct schedule_node
*node
;
72 bool write_after_read
;
75 /* When walking the instructions in reverse, we need to swap before/after in
78 enum direction
{ F
, R
};
80 struct schedule_state
{
81 const struct v3d_device_info
*devinfo
;
82 struct schedule_node
*last_r
[6];
83 struct schedule_node
*last_rf
[64];
84 struct schedule_node
*last_sf
;
85 struct schedule_node
*last_vpm_read
;
86 struct schedule_node
*last_tmu_write
;
87 struct schedule_node
*last_tmu_config
;
88 struct schedule_node
*last_tlb
;
89 struct schedule_node
*last_vpm
;
90 struct schedule_node
*last_unif
;
91 struct schedule_node
*last_rtop
;
93 /* Estimated cycle when the current instruction would start. */
98 add_dep(struct schedule_state
*state
,
99 struct schedule_node
*before
,
100 struct schedule_node
*after
,
103 bool write_after_read
= !write
&& state
->dir
== R
;
105 if (!before
|| !after
)
108 assert(before
!= after
);
110 if (state
->dir
== R
) {
111 struct schedule_node
*t
= before
;
116 for (int i
= 0; i
< before
->child_count
; i
++) {
117 if (before
->children
[i
].node
== after
&&
118 (before
->children
[i
].write_after_read
== write_after_read
)) {
123 if (before
->child_array_size
<= before
->child_count
) {
124 before
->child_array_size
= MAX2(before
->child_array_size
* 2, 16);
125 before
->children
= reralloc(before
, before
->children
,
126 struct schedule_node_child
,
127 before
->child_array_size
);
130 before
->children
[before
->child_count
].node
= after
;
131 before
->children
[before
->child_count
].write_after_read
=
133 before
->child_count
++;
134 after
->parent_count
++;
138 add_read_dep(struct schedule_state
*state
,
139 struct schedule_node
*before
,
140 struct schedule_node
*after
)
142 add_dep(state
, before
, after
, false);
146 add_write_dep(struct schedule_state
*state
,
147 struct schedule_node
**before
,
148 struct schedule_node
*after
)
150 add_dep(state
, *before
, after
, true);
155 qpu_inst_is_tlb(const struct v3d_qpu_instr
*inst
)
157 if (inst
->type
!= V3D_QPU_INSTR_TYPE_ALU
)
160 if (inst
->alu
.add
.magic_write
&&
161 (inst
->alu
.add
.waddr
== V3D_QPU_WADDR_TLB
||
162 inst
->alu
.add
.waddr
== V3D_QPU_WADDR_TLBU
))
165 if (inst
->alu
.mul
.magic_write
&&
166 (inst
->alu
.mul
.waddr
== V3D_QPU_WADDR_TLB
||
167 inst
->alu
.mul
.waddr
== V3D_QPU_WADDR_TLBU
))
174 process_mux_deps(struct schedule_state
*state
, struct schedule_node
*n
,
175 enum v3d_qpu_mux mux
)
179 add_read_dep(state
, state
->last_rf
[n
->inst
->qpu
.raddr_a
], n
);
182 add_read_dep(state
, state
->last_rf
[n
->inst
->qpu
.raddr_b
], n
);
185 add_read_dep(state
, state
->last_r
[mux
- V3D_QPU_MUX_R0
], n
);
192 process_waddr_deps(struct schedule_state
*state
, struct schedule_node
*n
,
193 uint32_t waddr
, bool magic
)
196 add_write_dep(state
, &state
->last_rf
[waddr
], n
);
197 } else if (v3d_qpu_magic_waddr_is_tmu(waddr
)) {
198 add_write_dep(state
, &state
->last_tmu_write
, n
);
200 case V3D_QPU_WADDR_TMUS
:
201 case V3D_QPU_WADDR_TMUSCM
:
202 case V3D_QPU_WADDR_TMUSF
:
203 case V3D_QPU_WADDR_TMUSLOD
:
204 add_write_dep(state
, &state
->last_tmu_config
, n
);
209 } else if (v3d_qpu_magic_waddr_is_sfu(waddr
)) {
210 /* Handled by v3d_qpu_writes_r4() check. */
213 case V3D_QPU_WADDR_R0
:
214 case V3D_QPU_WADDR_R1
:
215 case V3D_QPU_WADDR_R2
:
217 &state
->last_r
[waddr
- V3D_QPU_WADDR_R0
],
220 case V3D_QPU_WADDR_R3
:
221 case V3D_QPU_WADDR_R4
:
222 case V3D_QPU_WADDR_R5
:
223 /* Handled by v3d_qpu_writes_r*() checks below. */
226 case V3D_QPU_WADDR_VPM
:
227 case V3D_QPU_WADDR_VPMU
:
228 add_write_dep(state
, &state
->last_vpm
, n
);
231 case V3D_QPU_WADDR_TLB
:
232 case V3D_QPU_WADDR_TLBU
:
233 add_write_dep(state
, &state
->last_tlb
, n
);
236 case V3D_QPU_WADDR_NOP
:
240 fprintf(stderr
, "Unknown waddr %d\n", waddr
);
247 process_cond_deps(struct schedule_state
*state
, struct schedule_node
*n
,
248 enum v3d_qpu_cond cond
)
250 if (cond
!= V3D_QPU_COND_NONE
)
251 add_read_dep(state
, state
->last_sf
, n
);
255 process_pf_deps(struct schedule_state
*state
, struct schedule_node
*n
,
258 if (pf
!= V3D_QPU_PF_NONE
)
259 add_write_dep(state
, &state
->last_sf
, n
);
263 process_uf_deps(struct schedule_state
*state
, struct schedule_node
*n
,
266 if (uf
!= V3D_QPU_UF_NONE
)
267 add_write_dep(state
, &state
->last_sf
, n
);
271 * Common code for dependencies that need to be tracked both forward and
274 * This is for things like "all reads of r4 have to happen between the r4
275 * writes that surround them".
278 calculate_deps(struct schedule_state
*state
, struct schedule_node
*n
)
280 const struct v3d_device_info
*devinfo
= state
->devinfo
;
281 struct qinst
*qinst
= n
->inst
;
282 struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
284 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
) {
285 if (inst
->branch
.cond
!= V3D_QPU_BRANCH_COND_ALWAYS
)
286 add_read_dep(state
, state
->last_sf
, n
);
293 add_write_dep(state
, &state
->last_unif
, n
);
297 assert(inst
->type
== V3D_QPU_INSTR_TYPE_ALU
);
301 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 0)
302 process_mux_deps(state
, n
, inst
->alu
.add
.a
);
303 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 1)
304 process_mux_deps(state
, n
, inst
->alu
.add
.b
);
306 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 0)
307 process_mux_deps(state
, n
, inst
->alu
.mul
.a
);
308 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 1)
309 process_mux_deps(state
, n
, inst
->alu
.mul
.b
);
311 switch (inst
->alu
.add
.op
) {
312 case V3D_QPU_A_VPMSETUP
:
313 /* Could distinguish read/write by unpacking the uniform. */
314 add_write_dep(state
, &state
->last_vpm
, n
);
315 add_write_dep(state
, &state
->last_vpm_read
, n
);
318 case V3D_QPU_A_STVPMV
:
319 case V3D_QPU_A_STVPMD
:
320 case V3D_QPU_A_STVPMP
:
321 add_write_dep(state
, &state
->last_vpm
, n
);
324 case V3D_QPU_A_VPMWT
:
325 add_read_dep(state
, state
->last_vpm
, n
);
329 add_read_dep(state
, state
->last_tlb
, n
);
332 case V3D_QPU_A_SETMSF
:
333 case V3D_QPU_A_SETREVF
:
334 add_write_dep(state
, &state
->last_tlb
, n
);
337 case V3D_QPU_A_FLAPUSH
:
338 case V3D_QPU_A_FLBPUSH
:
340 case V3D_QPU_A_VFLNA
:
342 case V3D_QPU_A_VFLNB
:
343 add_read_dep(state
, state
->last_sf
, n
);
346 case V3D_QPU_A_FLBPOP
:
347 add_write_dep(state
, &state
->last_sf
, n
);
354 switch (inst
->alu
.mul
.op
) {
355 case V3D_QPU_M_MULTOP
:
356 case V3D_QPU_M_UMUL24
:
357 /* MULTOP sets rtop, and UMUL24 implicitly reads rtop and
358 * resets it to 0. We could possibly reorder umul24s relative
359 * to each other, but for now just keep all the MUL parts in
362 add_write_dep(state
, &state
->last_rtop
, n
);
368 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
369 process_waddr_deps(state
, n
, inst
->alu
.add
.waddr
,
370 inst
->alu
.add
.magic_write
);
372 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
373 process_waddr_deps(state
, n
, inst
->alu
.mul
.waddr
,
374 inst
->alu
.mul
.magic_write
);
376 if (v3d_qpu_sig_writes_address(devinfo
, &inst
->sig
)) {
377 process_waddr_deps(state
, n
, inst
->sig_addr
,
381 if (v3d_qpu_writes_r3(devinfo
, inst
))
382 add_write_dep(state
, &state
->last_r
[3], n
);
383 if (v3d_qpu_writes_r4(devinfo
, inst
))
384 add_write_dep(state
, &state
->last_r
[4], n
);
385 if (v3d_qpu_writes_r5(devinfo
, inst
))
386 add_write_dep(state
, &state
->last_r
[5], n
);
388 if (inst
->sig
.thrsw
) {
389 /* All accumulator contents and flags are undefined after the
392 for (int i
= 0; i
< ARRAY_SIZE(state
->last_r
); i
++)
393 add_write_dep(state
, &state
->last_r
[i
], n
);
394 add_write_dep(state
, &state
->last_sf
, n
);
396 /* Scoreboard-locking operations have to stay after the last
399 add_write_dep(state
, &state
->last_tlb
, n
);
401 add_write_dep(state
, &state
->last_tmu_write
, n
);
402 add_write_dep(state
, &state
->last_tmu_config
, n
);
405 if (inst
->sig
.ldtmu
) {
406 /* TMU loads are coming from a FIFO, so ordering is important.
408 add_write_dep(state
, &state
->last_tmu_write
, n
);
411 if (inst
->sig
.wrtmuc
)
412 add_write_dep(state
, &state
->last_tmu_config
, n
);
414 if (inst
->sig
.ldtlb
| inst
->sig
.ldtlbu
)
415 add_read_dep(state
, state
->last_tlb
, n
);
418 add_write_dep(state
, &state
->last_vpm_read
, n
);
420 /* inst->sig.ldunif or sideband uniform read */
421 if (qinst
->uniform
!= ~0)
422 add_write_dep(state
, &state
->last_unif
, n
);
424 process_cond_deps(state
, n
, inst
->flags
.ac
);
425 process_cond_deps(state
, n
, inst
->flags
.mc
);
426 process_pf_deps(state
, n
, inst
->flags
.apf
);
427 process_pf_deps(state
, n
, inst
->flags
.mpf
);
428 process_uf_deps(state
, n
, inst
->flags
.auf
);
429 process_uf_deps(state
, n
, inst
->flags
.muf
);
433 calculate_forward_deps(struct v3d_compile
*c
, struct list_head
*schedule_list
)
435 struct schedule_state state
;
437 memset(&state
, 0, sizeof(state
));
438 state
.devinfo
= c
->devinfo
;
441 list_for_each_entry(struct schedule_node
, node
, schedule_list
, link
)
442 calculate_deps(&state
, node
);
446 calculate_reverse_deps(struct v3d_compile
*c
, struct list_head
*schedule_list
)
448 struct list_head
*node
;
449 struct schedule_state state
;
451 memset(&state
, 0, sizeof(state
));
452 state
.devinfo
= c
->devinfo
;
455 for (node
= schedule_list
->prev
; schedule_list
!= node
; node
= node
->prev
) {
456 calculate_deps(&state
, (struct schedule_node
*)node
);
460 struct choose_scoreboard
{
462 int last_sfu_write_tick
;
463 int last_ldvary_tick
;
464 int last_uniforms_reset_tick
;
465 uint32_t last_waddr_add
, last_waddr_mul
;
470 mux_reads_too_soon(struct choose_scoreboard
*scoreboard
,
471 const struct v3d_qpu_instr
*inst
, enum v3d_qpu_mux mux
)
475 if (scoreboard
->last_waddr_add
== inst
->raddr_a
||
476 scoreboard
->last_waddr_mul
== inst
->raddr_a
) {
482 if (scoreboard
->last_waddr_add
== inst
->raddr_b
||
483 scoreboard
->last_waddr_mul
== inst
->raddr_b
) {
489 if (scoreboard
->tick
- scoreboard
->last_sfu_write_tick
<= 2)
494 if (scoreboard
->tick
- scoreboard
->last_ldvary_tick
<= 1)
505 reads_too_soon_after_write(struct choose_scoreboard
*scoreboard
,
508 const struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
510 /* XXX: Branching off of raddr. */
511 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
)
514 assert(inst
->type
== V3D_QPU_INSTR_TYPE_ALU
);
516 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
517 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 0 &&
518 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.add
.a
)) {
521 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 1 &&
522 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.add
.b
)) {
527 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
528 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 0 &&
529 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.mul
.a
)) {
532 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 1 &&
533 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.mul
.b
)) {
544 writes_too_soon_after_write(const struct v3d_device_info
*devinfo
,
545 struct choose_scoreboard
*scoreboard
,
548 const struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
550 /* Don't schedule any other r4 write too soon after an SFU write.
551 * This would normally be prevented by dependency tracking, but might
552 * occur if a dead SFU computation makes it to scheduling.
554 if (scoreboard
->tick
- scoreboard
->last_sfu_write_tick
< 2 &&
555 v3d_qpu_writes_r4(devinfo
, inst
))
562 pixel_scoreboard_too_soon(struct choose_scoreboard
*scoreboard
,
563 const struct v3d_qpu_instr
*inst
)
565 return (scoreboard
->tick
== 0 && qpu_inst_is_tlb(inst
));
569 get_instruction_priority(const struct v3d_qpu_instr
*inst
)
571 uint32_t baseline_score
;
572 uint32_t next_score
= 0;
574 /* Schedule TLB operations as late as possible, to get more
575 * parallelism between shaders.
577 if (qpu_inst_is_tlb(inst
))
581 /* Schedule texture read results collection late to hide latency. */
586 /* Default score for things that aren't otherwise special. */
587 baseline_score
= next_score
;
590 /* Schedule texture read setup early to hide their latency better. */
591 if (inst
->type
== V3D_QPU_INSTR_TYPE_ALU
&&
592 ((inst
->alu
.add
.magic_write
&&
593 v3d_qpu_magic_waddr_is_tmu(inst
->alu
.add
.waddr
)) ||
594 (inst
->alu
.mul
.magic_write
&&
595 v3d_qpu_magic_waddr_is_tmu(inst
->alu
.mul
.waddr
)))) {
600 return baseline_score
;
604 qpu_magic_waddr_is_periph(enum v3d_qpu_waddr waddr
)
606 return (v3d_qpu_magic_waddr_is_tmu(waddr
) ||
607 v3d_qpu_magic_waddr_is_sfu(waddr
) ||
608 v3d_qpu_magic_waddr_is_tlb(waddr
) ||
609 v3d_qpu_magic_waddr_is_vpm(waddr
) ||
610 v3d_qpu_magic_waddr_is_tsy(waddr
));
614 qpu_accesses_peripheral(const struct v3d_qpu_instr
*inst
)
616 if (v3d_qpu_uses_vpm(inst
))
619 if (inst
->type
== V3D_QPU_INSTR_TYPE_ALU
) {
620 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
&&
621 inst
->alu
.add
.magic_write
&&
622 qpu_magic_waddr_is_periph(inst
->alu
.add
.waddr
)) {
626 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
&&
627 inst
->alu
.mul
.magic_write
&&
628 qpu_magic_waddr_is_periph(inst
->alu
.mul
.waddr
)) {
633 return (inst
->sig
.ldvpm
||
641 qpu_merge_inst(const struct v3d_device_info
*devinfo
,
642 struct v3d_qpu_instr
*result
,
643 const struct v3d_qpu_instr
*a
,
644 const struct v3d_qpu_instr
*b
)
646 if (a
->type
!= V3D_QPU_INSTR_TYPE_ALU
||
647 b
->type
!= V3D_QPU_INSTR_TYPE_ALU
) {
651 /* Can't do more than one peripheral access in an instruction.
653 * XXX: V3D 4.1 allows TMU read along with a VPM read or write, and
654 * WRTMUC with a TMU magic register write (other than tmuc).
656 if (qpu_accesses_peripheral(a
) && qpu_accesses_peripheral(b
))
659 struct v3d_qpu_instr merge
= *a
;
661 if (b
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
662 if (a
->alu
.add
.op
!= V3D_QPU_A_NOP
)
664 merge
.alu
.add
= b
->alu
.add
;
666 merge
.flags
.ac
= b
->flags
.ac
;
667 merge
.flags
.apf
= b
->flags
.apf
;
668 merge
.flags
.auf
= b
->flags
.auf
;
671 if (b
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
672 if (a
->alu
.mul
.op
!= V3D_QPU_M_NOP
)
674 merge
.alu
.mul
= b
->alu
.mul
;
676 merge
.flags
.mc
= b
->flags
.mc
;
677 merge
.flags
.mpf
= b
->flags
.mpf
;
678 merge
.flags
.muf
= b
->flags
.muf
;
681 if (v3d_qpu_uses_mux(b
, V3D_QPU_MUX_A
)) {
682 if (v3d_qpu_uses_mux(a
, V3D_QPU_MUX_A
) &&
683 a
->raddr_a
!= b
->raddr_a
) {
686 merge
.raddr_a
= b
->raddr_a
;
689 if (v3d_qpu_uses_mux(b
, V3D_QPU_MUX_B
)) {
690 if (v3d_qpu_uses_mux(a
, V3D_QPU_MUX_B
) &&
691 a
->raddr_b
!= b
->raddr_b
) {
694 merge
.raddr_b
= b
->raddr_b
;
697 merge
.sig
.thrsw
|= b
->sig
.thrsw
;
698 merge
.sig
.ldunif
|= b
->sig
.ldunif
;
699 merge
.sig
.ldunifrf
|= b
->sig
.ldunifrf
;
700 merge
.sig
.ldunifa
|= b
->sig
.ldunifa
;
701 merge
.sig
.ldunifarf
|= b
->sig
.ldunifarf
;
702 merge
.sig
.ldtmu
|= b
->sig
.ldtmu
;
703 merge
.sig
.ldvary
|= b
->sig
.ldvary
;
704 merge
.sig
.ldvpm
|= b
->sig
.ldvpm
;
705 merge
.sig
.small_imm
|= b
->sig
.small_imm
;
706 merge
.sig
.ldtlb
|= b
->sig
.ldtlb
;
707 merge
.sig
.ldtlbu
|= b
->sig
.ldtlbu
;
708 merge
.sig
.ucb
|= b
->sig
.ucb
;
709 merge
.sig
.rotate
|= b
->sig
.rotate
;
710 merge
.sig
.wrtmuc
|= b
->sig
.wrtmuc
;
712 if (v3d_qpu_sig_writes_address(devinfo
, &a
->sig
) &&
713 v3d_qpu_sig_writes_address(devinfo
, &b
->sig
))
715 merge
.sig_addr
|= b
->sig_addr
;
716 merge
.sig_magic
|= b
->sig_magic
;
719 bool ok
= v3d_qpu_instr_pack(devinfo
, &merge
, &packed
);
722 /* No modifying the real instructions on failure. */
723 assert(ok
|| (a
!= result
&& b
!= result
));
728 static struct schedule_node
*
729 choose_instruction_to_schedule(const struct v3d_device_info
*devinfo
,
730 struct choose_scoreboard
*scoreboard
,
731 struct list_head
*schedule_list
,
732 struct schedule_node
*prev_inst
)
734 struct schedule_node
*chosen
= NULL
;
737 /* Don't pair up anything with a thread switch signal -- emit_thrsw()
738 * will handle pairing it along with filling the delay slots.
741 if (prev_inst
->inst
->qpu
.sig
.thrsw
)
745 list_for_each_entry(struct schedule_node
, n
, schedule_list
, link
) {
746 const struct v3d_qpu_instr
*inst
= &n
->inst
->qpu
;
748 /* Don't choose the branch instruction until it's the last one
749 * left. We'll move it up to fit its delay slots after we
752 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
&&
753 !list_is_singular(schedule_list
)) {
757 /* "An instruction must not read from a location in physical
758 * regfile A or B that was written to by the previous
761 if (reads_too_soon_after_write(scoreboard
, n
->inst
))
764 if (writes_too_soon_after_write(devinfo
, scoreboard
, n
->inst
))
767 /* "A scoreboard wait must not occur in the first two
768 * instructions of a fragment shader. This is either the
769 * explicit Wait for Scoreboard signal or an implicit wait
770 * with the first tile-buffer read or write instruction."
772 if (pixel_scoreboard_too_soon(scoreboard
, inst
))
775 /* ldunif and ldvary both write r5, but ldunif does so a tick
776 * sooner. If the ldvary's r5 wasn't used, then ldunif might
777 * otherwise get scheduled so ldunif and ldvary try to update
778 * r5 in the same tick.
780 if ((inst
->sig
.ldunif
|| inst
->sig
.ldunifa
) &&
781 scoreboard
->tick
== scoreboard
->last_ldvary_tick
+ 1) {
785 /* If we're trying to pair with another instruction, check
786 * that they're compatible.
789 /* Don't pair up a thread switch signal -- we'll
790 * handle pairing it when we pick it on its own.
795 if (prev_inst
->inst
->uniform
!= -1 &&
796 n
->inst
->uniform
!= -1)
799 /* Don't merge in something that will lock the TLB.
800 * Hopwefully what we have in inst will release some
801 * other instructions, allowing us to delay the
802 * TLB-locking instruction until later.
804 if (!scoreboard
->tlb_locked
&& qpu_inst_is_tlb(inst
))
807 struct v3d_qpu_instr merged_inst
;
808 if (!qpu_merge_inst(devinfo
, &merged_inst
,
809 &prev_inst
->inst
->qpu
, inst
)) {
814 int prio
= get_instruction_priority(inst
);
816 /* Found a valid instruction. If nothing better comes along,
825 if (prio
> chosen_prio
) {
828 } else if (prio
< chosen_prio
) {
832 if (n
->delay
> chosen
->delay
) {
835 } else if (n
->delay
< chosen
->delay
) {
844 update_scoreboard_for_magic_waddr(struct choose_scoreboard
*scoreboard
,
845 enum v3d_qpu_waddr waddr
)
847 if (v3d_qpu_magic_waddr_is_sfu(waddr
))
848 scoreboard
->last_sfu_write_tick
= scoreboard
->tick
;
852 update_scoreboard_for_chosen(struct choose_scoreboard
*scoreboard
,
853 const struct v3d_qpu_instr
*inst
)
855 scoreboard
->last_waddr_add
= ~0;
856 scoreboard
->last_waddr_mul
= ~0;
858 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
)
861 assert(inst
->type
== V3D_QPU_INSTR_TYPE_ALU
);
863 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
864 if (inst
->alu
.add
.magic_write
) {
865 update_scoreboard_for_magic_waddr(scoreboard
,
866 inst
->alu
.add
.waddr
);
868 scoreboard
->last_waddr_add
= inst
->alu
.add
.waddr
;
872 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
873 if (inst
->alu
.mul
.magic_write
) {
874 update_scoreboard_for_magic_waddr(scoreboard
,
875 inst
->alu
.mul
.waddr
);
877 scoreboard
->last_waddr_mul
= inst
->alu
.mul
.waddr
;
881 if (inst
->sig
.ldvary
)
882 scoreboard
->last_ldvary_tick
= scoreboard
->tick
;
884 if (qpu_inst_is_tlb(inst
))
885 scoreboard
->tlb_locked
= true;
889 dump_state(const struct v3d_device_info
*devinfo
,
890 struct list_head
*schedule_list
)
892 list_for_each_entry(struct schedule_node
, n
, schedule_list
, link
) {
893 fprintf(stderr
, " t=%4d: ", n
->unblocked_time
);
894 v3d_qpu_dump(devinfo
, &n
->inst
->qpu
);
895 fprintf(stderr
, "\n");
897 for (int i
= 0; i
< n
->child_count
; i
++) {
898 struct schedule_node
*child
= n
->children
[i
].node
;
902 fprintf(stderr
, " - ");
903 v3d_qpu_dump(devinfo
, &child
->inst
->qpu
);
904 fprintf(stderr
, " (%d parents, %c)\n",
906 n
->children
[i
].write_after_read
? 'w' : 'r');
911 static uint32_t magic_waddr_latency(enum v3d_qpu_waddr waddr
,
912 const struct v3d_qpu_instr
*after
)
914 /* Apply some huge latency between texture fetch requests and getting
915 * their results back.
917 * FIXME: This is actually pretty bogus. If we do:
926 * we count that as worse than
935 * because we associate the first load_tmu0 with the *second* tmu0_s.
937 if (v3d_qpu_magic_waddr_is_tmu(waddr
) && after
->sig
.ldtmu
)
940 /* Assume that anything depending on us is consuming the SFU result. */
941 if (v3d_qpu_magic_waddr_is_sfu(waddr
))
948 instruction_latency(struct schedule_node
*before
, struct schedule_node
*after
)
950 const struct v3d_qpu_instr
*before_inst
= &before
->inst
->qpu
;
951 const struct v3d_qpu_instr
*after_inst
= &after
->inst
->qpu
;
952 uint32_t latency
= 1;
954 if (before_inst
->type
!= V3D_QPU_INSTR_TYPE_ALU
||
955 after_inst
->type
!= V3D_QPU_INSTR_TYPE_ALU
)
958 if (before_inst
->alu
.add
.magic_write
) {
959 latency
= MAX2(latency
,
960 magic_waddr_latency(before_inst
->alu
.add
.waddr
,
964 if (before_inst
->alu
.mul
.magic_write
) {
965 latency
= MAX2(latency
,
966 magic_waddr_latency(before_inst
->alu
.mul
.waddr
,
973 /** Recursive computation of the delay member of a node. */
975 compute_delay(struct schedule_node
*n
)
977 if (!n
->child_count
) {
980 for (int i
= 0; i
< n
->child_count
; i
++) {
981 if (!n
->children
[i
].node
->delay
)
982 compute_delay(n
->children
[i
].node
);
983 n
->delay
= MAX2(n
->delay
,
984 n
->children
[i
].node
->delay
+
985 instruction_latency(n
, n
->children
[i
].node
));
991 mark_instruction_scheduled(struct list_head
*schedule_list
,
993 struct schedule_node
*node
,
999 for (int i
= node
->child_count
- 1; i
>= 0; i
--) {
1000 struct schedule_node
*child
=
1001 node
->children
[i
].node
;
1006 if (war_only
&& !node
->children
[i
].write_after_read
)
1009 /* If the requirement is only that the node not appear before
1010 * the last read of its destination, then it can be scheduled
1011 * immediately after (or paired with!) the thing reading the
1014 uint32_t latency
= 0;
1016 latency
= instruction_latency(node
,
1017 node
->children
[i
].node
);
1020 child
->unblocked_time
= MAX2(child
->unblocked_time
,
1022 child
->parent_count
--;
1023 if (child
->parent_count
== 0)
1024 list_add(&child
->link
, schedule_list
);
1026 node
->children
[i
].node
= NULL
;
1031 insert_scheduled_instruction(struct v3d_compile
*c
,
1032 struct qblock
*block
,
1033 struct choose_scoreboard
*scoreboard
,
1036 list_addtail(&inst
->link
, &block
->instructions
);
1038 update_scoreboard_for_chosen(scoreboard
, &inst
->qpu
);
1039 c
->qpu_inst_count
++;
1043 static struct qinst
*
1046 struct qreg undef
= { QFILE_NULL
, 0 };
1047 struct qinst
*qinst
= vir_add_inst(V3D_QPU_A_NOP
, undef
, undef
, undef
);
1053 emit_nop(struct v3d_compile
*c
, struct qblock
*block
,
1054 struct choose_scoreboard
*scoreboard
)
1056 insert_scheduled_instruction(c
, block
, scoreboard
, vir_nop());
1060 qpu_instruction_valid_in_thrend_slot(struct v3d_compile
*c
,
1061 const struct qinst
*qinst
, int slot
)
1063 const struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
1065 /* Only TLB Z writes are prohibited in the last slot, but we don't
1066 * have those flagged so prohibit all TLB ops for now.
1068 if (slot
== 2 && qpu_inst_is_tlb(inst
))
1071 if (slot
> 0 && qinst
->uniform
!= ~0)
1074 if (v3d_qpu_uses_vpm(inst
))
1077 if (inst
->sig
.ldvary
)
1080 if (inst
->type
== V3D_QPU_INSTR_TYPE_ALU
) {
1081 /* No writing physical registers at the end. */
1082 if (!inst
->alu
.add
.magic_write
||
1083 !inst
->alu
.mul
.magic_write
) {
1087 if (c
->devinfo
->ver
< 40 && inst
->alu
.add
.op
== V3D_QPU_A_SETMSF
)
1090 /* RF0-2 might be overwritten during the delay slots by
1091 * fragment shader setup.
1093 if (inst
->raddr_a
< 3 &&
1094 (inst
->alu
.add
.a
== V3D_QPU_MUX_A
||
1095 inst
->alu
.add
.b
== V3D_QPU_MUX_A
||
1096 inst
->alu
.mul
.a
== V3D_QPU_MUX_A
||
1097 inst
->alu
.mul
.b
== V3D_QPU_MUX_A
)) {
1101 if (inst
->raddr_b
< 3 &&
1102 !inst
->sig
.small_imm
&&
1103 (inst
->alu
.add
.a
== V3D_QPU_MUX_B
||
1104 inst
->alu
.add
.b
== V3D_QPU_MUX_B
||
1105 inst
->alu
.mul
.a
== V3D_QPU_MUX_B
||
1106 inst
->alu
.mul
.b
== V3D_QPU_MUX_B
)) {
1115 valid_thrsw_sequence(struct v3d_compile
*c
,
1116 struct qinst
*qinst
, int instructions_in_sequence
,
1119 for (int slot
= 0; slot
< instructions_in_sequence
; slot
++) {
1120 /* No scheduling SFU when the result would land in the other
1121 * thread. The simulator complains for safety, though it
1122 * would only occur for dead code in our case.
1125 qinst
->qpu
.type
== V3D_QPU_INSTR_TYPE_ALU
&&
1126 (v3d_qpu_magic_waddr_is_sfu(qinst
->qpu
.alu
.add
.waddr
) ||
1127 v3d_qpu_magic_waddr_is_sfu(qinst
->qpu
.alu
.mul
.waddr
))) {
1131 if (slot
> 0 && qinst
->qpu
.sig
.ldvary
)
1135 !qpu_instruction_valid_in_thrend_slot(c
, qinst
, slot
)) {
1139 /* Note that the list is circular, so we can only do this up
1140 * to instructions_in_sequence.
1142 qinst
= (struct qinst
*)qinst
->link
.next
;
1149 * Emits a THRSW signal in the stream, trying to move it up to pair with
1150 * another instruction.
1153 emit_thrsw(struct v3d_compile
*c
,
1154 struct qblock
*block
,
1155 struct choose_scoreboard
*scoreboard
,
1161 /* There should be nothing in a thrsw inst being scheduled other than
1164 assert(inst
->qpu
.type
== V3D_QPU_INSTR_TYPE_ALU
);
1165 assert(inst
->qpu
.alu
.add
.op
== V3D_QPU_A_NOP
);
1166 assert(inst
->qpu
.alu
.mul
.op
== V3D_QPU_M_NOP
);
1168 /* Find how far back into previous instructions we can put the THRSW. */
1169 int slots_filled
= 0;
1170 struct qinst
*merge_inst
= NULL
;
1171 vir_for_each_inst_rev(prev_inst
, block
) {
1172 struct v3d_qpu_sig sig
= prev_inst
->qpu
.sig
;
1174 uint32_t packed_sig
;
1176 if (!v3d_qpu_sig_pack(c
->devinfo
, &sig
, &packed_sig
))
1179 if (!valid_thrsw_sequence(c
, prev_inst
, slots_filled
+ 1,
1184 merge_inst
= prev_inst
;
1185 if (++slots_filled
== 3)
1189 bool needs_free
= false;
1191 merge_inst
->qpu
.sig
.thrsw
= true;
1194 insert_scheduled_instruction(c
, block
, scoreboard
, inst
);
1200 /* Insert any extra delay slot NOPs we need. */
1201 for (int i
= 0; i
< 3 - slots_filled
; i
++) {
1202 emit_nop(c
, block
, scoreboard
);
1206 /* If we're emitting the last THRSW (other than program end), then
1207 * signal that to the HW by emitting two THRSWs in a row.
1209 if (inst
->is_last_thrsw
) {
1210 struct qinst
*second_inst
=
1211 (struct qinst
*)merge_inst
->link
.next
;
1212 second_inst
->qpu
.sig
.thrsw
= true;
1215 /* If we put our THRSW into another instruction, free up the
1216 * instruction that didn't end up scheduled into the list.
1225 schedule_instructions(struct v3d_compile
*c
,
1226 struct choose_scoreboard
*scoreboard
,
1227 struct qblock
*block
,
1228 struct list_head
*schedule_list
,
1229 enum quniform_contents
*orig_uniform_contents
,
1230 uint32_t *orig_uniform_data
,
1231 uint32_t *next_uniform
)
1233 const struct v3d_device_info
*devinfo
= c
->devinfo
;
1237 fprintf(stderr
, "initial deps:\n");
1238 dump_state(devinfo
, schedule_list
);
1239 fprintf(stderr
, "\n");
1242 /* Remove non-DAG heads from the list. */
1243 list_for_each_entry_safe(struct schedule_node
, n
, schedule_list
, link
) {
1244 if (n
->parent_count
!= 0)
1248 while (!list_empty(schedule_list
)) {
1249 struct schedule_node
*chosen
=
1250 choose_instruction_to_schedule(devinfo
,
1254 struct schedule_node
*merge
= NULL
;
1256 /* If there are no valid instructions to schedule, drop a NOP
1259 struct qinst
*qinst
= chosen
? chosen
->inst
: vir_nop();
1260 struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
1263 fprintf(stderr
, "t=%4d: current list:\n",
1265 dump_state(devinfo
, schedule_list
);
1266 fprintf(stderr
, "t=%4d: chose: ", time
);
1267 v3d_qpu_dump(devinfo
, inst
);
1268 fprintf(stderr
, "\n");
1271 /* We can't mark_instruction_scheduled() the chosen inst until
1272 * we're done identifying instructions to merge, so put the
1273 * merged instructions on a list for a moment.
1275 struct list_head merged_list
;
1276 list_inithead(&merged_list
);
1278 /* Schedule this instruction onto the QPU list. Also try to
1279 * find an instruction to pair with it.
1282 time
= MAX2(chosen
->unblocked_time
, time
);
1283 list_del(&chosen
->link
);
1284 mark_instruction_scheduled(schedule_list
, time
,
1288 choose_instruction_to_schedule(devinfo
,
1292 time
= MAX2(merge
->unblocked_time
, time
);
1293 list_del(&merge
->link
);
1294 list_addtail(&merge
->link
, &merged_list
);
1295 (void)qpu_merge_inst(devinfo
, inst
,
1296 inst
, &merge
->inst
->qpu
);
1297 if (merge
->inst
->uniform
!= -1) {
1298 chosen
->inst
->uniform
=
1299 merge
->inst
->uniform
;
1303 fprintf(stderr
, "t=%4d: merging: ",
1305 v3d_qpu_dump(devinfo
, &merge
->inst
->qpu
);
1306 fprintf(stderr
, "\n");
1307 fprintf(stderr
, " result: ");
1308 v3d_qpu_dump(devinfo
, inst
);
1309 fprintf(stderr
, "\n");
1314 /* Update the uniform index for the rewritten location --
1315 * branch target updating will still need to change
1316 * c->uniform_data[] using this index.
1318 if (qinst
->uniform
!= -1) {
1319 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
)
1320 block
->branch_uniform
= *next_uniform
;
1322 c
->uniform_data
[*next_uniform
] =
1323 orig_uniform_data
[qinst
->uniform
];
1324 c
->uniform_contents
[*next_uniform
] =
1325 orig_uniform_contents
[qinst
->uniform
];
1326 qinst
->uniform
= *next_uniform
;
1331 fprintf(stderr
, "\n");
1334 /* Now that we've scheduled a new instruction, some of its
1335 * children can be promoted to the list of instructions ready to
1336 * be scheduled. Update the children's unblocked time for this
1337 * DAG edge as we do so.
1339 mark_instruction_scheduled(schedule_list
, time
, chosen
, false);
1340 list_for_each_entry(struct schedule_node
, merge
, &merged_list
,
1342 mark_instruction_scheduled(schedule_list
, time
, merge
,
1345 /* The merged VIR instruction doesn't get re-added to the
1346 * block, so free it now.
1351 if (inst
->sig
.thrsw
) {
1352 time
+= emit_thrsw(c
, block
, scoreboard
, qinst
, false);
1354 insert_scheduled_instruction(c
, block
,
1357 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
) {
1358 block
->branch_qpu_ip
= c
->qpu_inst_count
- 1;
1359 /* Fill the delay slots.
1361 * We should fill these with actual instructions,
1362 * instead, but that will probably need to be done
1363 * after this, once we know what the leading
1364 * instructions of the successors are (so we can
1365 * handle A/B register file write latency)
1367 for (int i
= 0; i
< 3; i
++)
1368 emit_nop(c
, block
, scoreboard
);
1377 qpu_schedule_instructions_block(struct v3d_compile
*c
,
1378 struct choose_scoreboard
*scoreboard
,
1379 struct qblock
*block
,
1380 enum quniform_contents
*orig_uniform_contents
,
1381 uint32_t *orig_uniform_data
,
1382 uint32_t *next_uniform
)
1384 void *mem_ctx
= ralloc_context(NULL
);
1385 struct list_head schedule_list
;
1387 list_inithead(&schedule_list
);
1389 /* Wrap each instruction in a scheduler structure. */
1390 while (!list_empty(&block
->instructions
)) {
1391 struct qinst
*qinst
= (struct qinst
*)block
->instructions
.next
;
1392 struct schedule_node
*n
=
1393 rzalloc(mem_ctx
, struct schedule_node
);
1397 list_del(&qinst
->link
);
1398 list_addtail(&n
->link
, &schedule_list
);
1401 calculate_forward_deps(c
, &schedule_list
);
1402 calculate_reverse_deps(c
, &schedule_list
);
1404 list_for_each_entry(struct schedule_node
, n
, &schedule_list
, link
) {
1408 uint32_t cycles
= schedule_instructions(c
, scoreboard
, block
,
1410 orig_uniform_contents
,
1414 ralloc_free(mem_ctx
);
1420 qpu_set_branch_targets(struct v3d_compile
*c
)
1422 vir_for_each_block(block
, c
) {
1423 /* The end block of the program has no branch. */
1424 if (!block
->successors
[0])
1427 /* If there was no branch instruction, then the successor
1428 * block must follow immediately after this one.
1430 if (block
->branch_qpu_ip
== ~0) {
1431 assert(block
->end_qpu_ip
+ 1 ==
1432 block
->successors
[0]->start_qpu_ip
);
1436 /* Walk back through the delay slots to find the branch
1439 struct list_head
*entry
= block
->instructions
.prev
;
1440 for (int i
= 0; i
< 3; i
++)
1441 entry
= entry
->prev
;
1442 struct qinst
*branch
= container_of(entry
, branch
, link
);
1443 assert(branch
->qpu
.type
== V3D_QPU_INSTR_TYPE_BRANCH
);
1445 /* Make sure that the if-we-don't-jump
1446 * successor was scheduled just after the
1449 assert(!block
->successors
[1] ||
1450 block
->successors
[1]->start_qpu_ip
==
1451 block
->branch_qpu_ip
+ 4);
1453 branch
->qpu
.branch
.offset
=
1454 ((block
->successors
[0]->start_qpu_ip
-
1455 (block
->branch_qpu_ip
+ 4)) *
1458 /* Set up the relative offset to jump in the
1461 * Use a temporary here, because
1462 * uniform_data[inst->uniform] may be shared
1463 * between multiple instructions.
1465 assert(c
->uniform_contents
[branch
->uniform
] == QUNIFORM_CONSTANT
);
1466 c
->uniform_data
[branch
->uniform
] =
1467 (block
->successors
[0]->start_uniform
-
1468 (block
->branch_uniform
+ 1)) * 4;
1473 v3d_qpu_schedule_instructions(struct v3d_compile
*c
)
1475 const struct v3d_device_info
*devinfo
= c
->devinfo
;
1476 struct qblock
*end_block
= list_last_entry(&c
->blocks
,
1477 struct qblock
, link
);
1479 /* We reorder the uniforms as we schedule instructions, so save the
1480 * old data off and replace it.
1482 uint32_t *uniform_data
= c
->uniform_data
;
1483 enum quniform_contents
*uniform_contents
= c
->uniform_contents
;
1484 c
->uniform_contents
= ralloc_array(c
, enum quniform_contents
,
1486 c
->uniform_data
= ralloc_array(c
, uint32_t, c
->num_uniforms
);
1487 c
->uniform_array_size
= c
->num_uniforms
;
1488 uint32_t next_uniform
= 0;
1490 struct choose_scoreboard scoreboard
;
1491 memset(&scoreboard
, 0, sizeof(scoreboard
));
1492 scoreboard
.last_waddr_add
= ~0;
1493 scoreboard
.last_waddr_mul
= ~0;
1494 scoreboard
.last_ldvary_tick
= -10;
1495 scoreboard
.last_sfu_write_tick
= -10;
1496 scoreboard
.last_uniforms_reset_tick
= -10;
1499 fprintf(stderr
, "Pre-schedule instructions\n");
1500 vir_for_each_block(block
, c
) {
1501 fprintf(stderr
, "BLOCK %d\n", block
->index
);
1502 list_for_each_entry(struct qinst
, qinst
,
1503 &block
->instructions
, link
) {
1504 v3d_qpu_dump(devinfo
, &qinst
->qpu
);
1505 fprintf(stderr
, "\n");
1508 fprintf(stderr
, "\n");
1511 uint32_t cycles
= 0;
1512 vir_for_each_block(block
, c
) {
1513 block
->start_qpu_ip
= c
->qpu_inst_count
;
1514 block
->branch_qpu_ip
= ~0;
1515 block
->start_uniform
= next_uniform
;
1517 cycles
+= qpu_schedule_instructions_block(c
,
1524 block
->end_qpu_ip
= c
->qpu_inst_count
- 1;
1527 /* Emit the program-end THRSW instruction. */;
1528 struct qinst
*thrsw
= vir_nop();
1529 thrsw
->qpu
.sig
.thrsw
= true;
1530 emit_thrsw(c
, end_block
, &scoreboard
, thrsw
, true);
1532 qpu_set_branch_targets(c
);
1534 assert(next_uniform
== c
->num_uniforms
);